Merge

src/cpu/sparc/vm/macroAssembler_sparc.cpp

@@ -3526,8 +3526,12 @@ void MacroAssembler::bang_stack_size(Register Rsize, Register Rtsp,
   delayed()->sub(Rtsp, Roffset, Rtsp);
 
   // Bang down shadow pages too.
-  // The -1 because we already subtracted 1 page.
-  for (int i = 0; i< StackShadowPages-1; i++) {
+  // At this point, (tmp-0) is the last address touched, so don't
+  // touch it again.  (It was touched as (tmp-pagesize) but then tmp
+  // was post-decremented.)  Skip this address by starting at i=1, and
+  // touch a few more pages below.  N.B.  It is important to touch all
+  // the way down to and including i=StackShadowPages.
+  for (int i = 1; i <= StackShadowPages; i++) {
     set((-i*offset)+STACK_BIAS, Rscratch);
     st(G0, Rtsp, Rscratch);
   }

src/cpu/sparc/vm/sparc.ad

@@ -2916,6 +2916,9 @@ enc_class Fast_Unlock(iRegP oop, iRegP box, o7RegP scratch, iRegP scratch2) %{
       __ bind(LSkip2);
     }
 
+    // We have no guarantee that on 64 bit the higher half of limit_reg is 0
+    __ signx(limit_reg);
+
     __ subcc(limit_reg, 1 * sizeof(jchar), chr1_reg);
     __ br(Assembler::equal, true, Assembler::pn, Ldone);
     __ delayed()->mov(O7, result_reg);  // result is difference in lengths
@@ -2973,6 +2976,9 @@ enc_class enc_String_Equals(o0RegP str1, o1RegP str2, g3RegI cnt, notemp_iRegI r
     Register  chr1_reg = result_reg;
     Register  chr2_reg =   tmp1_reg;
 
+    // We have no guarantee that on 64 bit the higher half of limit_reg is 0
+    __ signx(limit_reg);
+
     //check for alignment and position the pointers to the ends
     __ or3(str1_reg, str2_reg, chr1_reg);
     __ andcc(chr1_reg, 0x3, chr1_reg);

src/cpu/x86/vm/macroAssembler_x86.cpp

@@ -1381,8 +1381,12 @@ void MacroAssembler::bang_stack_size(Register size, Register tmp) {
   jcc(Assembler::greater, loop);
 
   // Bang down shadow pages too.
-  // The -1 because we already subtracted 1 page.
-  for (int i = 0; i< StackShadowPages-1; i++) {
+  // At this point, (tmp-0) is the last address touched, so don't
+  // touch it again.  (It was touched as (tmp-pagesize) but then tmp
+  // was post-decremented.)  Skip this address by starting at i=1, and
+  // touch a few more pages below.  N.B.  It is important to touch all
+  // the way down to and including i=StackShadowPages.
+  for (int i = 1; i <= StackShadowPages; i++) {
     // this could be any sized move but this is can be a debugging crumb
     // so the bigger the better.
     movptr(Address(tmp, (-i*os::vm_page_size())), size );

src/share/vm/asm/assembler.cpp

@@ -122,7 +122,7 @@ void AbstractAssembler::bind(Label& L) {
 void AbstractAssembler::generate_stack_overflow_check( int frame_size_in_bytes) {
   if (UseStackBanging) {
     // Each code entry causes one stack bang n pages down the stack where n
-    // is configurable by StackBangPages.  The setting depends on the maximum
+    // is configurable by StackShadowPages.  The setting depends on the maximum
     // depth of VM call stack or native before going back into java code,
     // since only java code can raise a stack overflow exception using the
     // stack banging mechanism.  The VM and native code does not detect stack

src/share/vm/c1/c1_LinearScan.cpp

@@ -1138,8 +1138,10 @@ IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) {
         }
       }
     }
-
-  } else if (opr_type != T_LONG) {
+    // We want to sometimes use logical operations on pointers, in particular in GC barriers.
+    // Since 64bit logical operations do not current support operands on stack, we have to make sure
+    // T_OBJECT doesn't get spilled along with T_LONG.
+  } else if (opr_type != T_LONG LP64_ONLY(&& opr_type != T_OBJECT)) {
     // integer instruction (note: long operands must always be in register)
     switch (op->code()) {
       case lir_cmp:

src/share/vm/opto/compile.cpp

@@ -848,6 +848,7 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
   }
 #endif
 
+  NOT_PRODUCT( verify_barriers(); )
   // Now that we know the size of all the monitors we can add a fixed slot
   // for the original deopt pc.
 
@@ -3018,12 +3019,17 @@ void Compile::final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc) {
           // Phi nodes shouldn't be moved. They would only match below if they
           // had the same control as the MathExactNode. The only time that
           // would happen is if the Phi is also an input to the MathExact
-          if (!out->is_Phi()) {
-            if (out->in(0) == NULL) {
-              out->set_req(0, non_throwing);
-            } else if (out->in(0) == ctrl) {
-              out->set_req(0, non_throwing);
-            }
+          //
+          // Cmp nodes shouldn't have control set at all.
+          if (out->is_Phi() ||
+              out->is_Cmp()) {
+            continue;
+          }
+
+          if (out->in(0) == NULL) {
+            out->set_req(0, non_throwing);
+          } else if (out->in(0) == ctrl) {
+            out->set_req(0, non_throwing);
           }
         }
       }
@@ -3368,6 +3374,72 @@ void Compile::verify_graph_edges(bool no_dead_code) {
     }
   }
 }
+
+// Verify GC barriers consistency
+// Currently supported:
+// - G1 pre-barriers (see GraphKit::g1_write_barrier_pre())
+void Compile::verify_barriers() {
+  if (UseG1GC) {
+    // Verify G1 pre-barriers
+    const int marking_offset = in_bytes(JavaThread::satb_mark_queue_offset() + PtrQueue::byte_offset_of_active());
+
+    ResourceArea *area = Thread::current()->resource_area();
+    Unique_Node_List visited(area);
+    Node_List worklist(area);
+    // We're going to walk control flow backwards starting from the Root
+    worklist.push(_root);
+    while (worklist.size() > 0) {
+      Node* x = worklist.pop();
+      if (x == NULL || x == top()) continue;
+      if (visited.member(x)) {
+        continue;
+      } else {
+        visited.push(x);
+      }
+
+      if (x->is_Region()) {
+        for (uint i = 1; i < x->req(); i++) {
+          worklist.push(x->in(i));
+        }
+      } else {
+        worklist.push(x->in(0));
+        // We are looking for the pattern:
+        //                            /->ThreadLocal
+        // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
+        //              \->ConI(0)
+        // We want to verify that the If and the LoadB have the same control
+        // See GraphKit::g1_write_barrier_pre()
+        if (x->is_If()) {
+          IfNode *iff = x->as_If();
+          if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
+            CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
+            if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
+                && cmp->in(1)->is_Load()) {
+              LoadNode* load = cmp->in(1)->as_Load();
+              if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
+                  && load->in(2)->in(3)->is_Con()
+                  && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
+
+                Node* if_ctrl = iff->in(0);
+                Node* load_ctrl = load->in(0);
+
+                if (if_ctrl != load_ctrl) {
+                  // Skip possible CProj->NeverBranch in infinite loops
+                  if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
+                      && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
+                    if_ctrl = if_ctrl->in(0)->in(0);
+                  }
+                }
+                assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+}
+
 #endif
 
 // The Compile object keeps track of failure reasons separately from the ciEnv.

src/share/vm/opto/compile.hpp

@@ -1148,6 +1148,9 @@ class Compile : public Phase {
   // graph is strongly connected from root in both directions.
   void verify_graph_edges(bool no_dead_code = false) PRODUCT_RETURN;
 
+  // Verify GC barrier patterns
+  void verify_barriers() PRODUCT_RETURN;
+
   // End-of-run dumps.
   static void print_statistics() PRODUCT_RETURN;
 

src/share/vm/opto/loopTransform.cpp

@@ -1964,7 +1964,7 @@ void PhaseIdealLoop::do_range_check( IdealLoopTree *loop, Node_List &old_new ) {
       // Find loads off the surviving projection; remove their control edge
       for (DUIterator_Fast imax, i = dp->fast_outs(imax); i < imax; i++) {
         Node* cd = dp->fast_out(i); // Control-dependent node
-        if( cd->is_Load() ) {   // Loads can now float around in the loop
+        if (cd->is_Load() && cd->depends_only_on_test()) {   // Loads can now float around in the loop
           // Allow the load to float around in the loop, or before it
           // but NOT before the pre-loop.
           _igvn.replace_input_of(cd, 0, ctrl); // ctrl, not NULL

src/share/vm/opto/memnode.hpp

@@ -204,6 +204,17 @@ public:
 protected:
   const Type* load_array_final_field(const TypeKlassPtr *tkls,
                                      ciKlass* klass) const;
+  // depends_only_on_test is almost always true, and needs to be almost always
+  // true to enable key hoisting & commoning optimizations.  However, for the
+  // special case of RawPtr loads from TLS top & end, and other loads performed by
+  // GC barriers, the control edge carries the dependence preventing hoisting past
+  // a Safepoint instead of the memory edge.  (An unfortunate consequence of having
+  // Safepoints not set Raw Memory; itself an unfortunate consequence of having Nodes
+  // which produce results (new raw memory state) inside of loops preventing all
+  // manner of other optimizations).  Basically, it's ugly but so is the alternative.
+  // See comment in macro.cpp, around line 125 expand_allocate_common().
+  virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; }
+
 };
 
 //------------------------------LoadBNode--------------------------------------
@@ -370,16 +381,6 @@ public:
   virtual uint ideal_reg() const { return Op_RegP; }
   virtual int store_Opcode() const { return Op_StoreP; }
   virtual BasicType memory_type() const { return T_ADDRESS; }
-  // depends_only_on_test is almost always true, and needs to be almost always
-  // true to enable key hoisting & commoning optimizations.  However, for the
-  // special case of RawPtr loads from TLS top & end, the control edge carries
-  // the dependence preventing hoisting past a Safepoint instead of the memory
-  // edge.  (An unfortunate consequence of having Safepoints not set Raw
-  // Memory; itself an unfortunate consequence of having Nodes which produce
-  // results (new raw memory state) inside of loops preventing all manner of
-  // other optimizations).  Basically, it's ugly but so is the alternative.
-  // See comment in macro.cpp, around line 125 expand_allocate_common().
-  virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; }
 };
 
 
@@ -393,16 +394,6 @@ public:
   virtual uint ideal_reg() const { return Op_RegN; }
   virtual int store_Opcode() const { return Op_StoreN; }
   virtual BasicType memory_type() const { return T_NARROWOOP; }
-  // depends_only_on_test is almost always true, and needs to be almost always
-  // true to enable key hoisting & commoning optimizations.  However, for the
-  // special case of RawPtr loads from TLS top & end, the control edge carries
-  // the dependence preventing hoisting past a Safepoint instead of the memory
-  // edge.  (An unfortunate consequence of having Safepoints not set Raw
-  // Memory; itself an unfortunate consequence of having Nodes which produce
-  // results (new raw memory state) inside of loops preventing all manner of
-  // other optimizations).  Basically, it's ugly but so is the alternative.
-  // See comment in macro.cpp, around line 125 expand_allocate_common().
-  virtual bool depends_only_on_test() const { return adr_type() != TypeRawPtr::BOTTOM; }
 };
 
 //------------------------------LoadKlassNode----------------------------------

test/compiler/intrinsics/mathexact/CompareTest.java

+/*
+ * Copyright (c) 2013, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ */
+
+/*
+ * @test
+ * @bug 8026722
+ * @summary Verify that the compare after addExact is a signed compare
+ * @compile CompareTest.java
+ * @run main CompareTest
+ *
+ */
+
+public class CompareTest {
+    public static long store = 0;
+    public static long addValue = 1231;
+
+    public static void main(String[] args) {
+        for (int i = 0; i < 20000; ++i) {
+            runTest(i, i);
+            runTest(i-1, i);
+        }
+    }
+
+    public static long create(long value, int v) {
+        if ((value | v) == 0) {
+            return 0;
+        }
+
+        // C2 turned this test into unsigned test when a control edge was set on the Cmp
+        if (value < -31557014167219200L || value > 31556889864403199L) {
+            throw new RuntimeException("error");
+        }
+
+        return value;
+    }
+
+    public static void runTest(long value, int value2) {
+        long res = Math.addExact(value, addValue);
+        store = create(res, Math.floorMod(value2, 100000));
+    }
+}

test/compiler/intrinsics/stringequals/TestStringEqualsBadLength.java

+/*
+ * Copyright (c) 2013, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ */
+
+/*
+ * @test
+ * @bug 8027445
+ * @summary String.equals() may be called with a length whose upper bits are not cleared
+ * @run main/othervm -XX:-UseOnStackReplacement -XX:-BackgroundCompilation TestStringEqualsBadLength
+ *
+ */
+
+import java.util.Arrays;
+
+public class TestStringEqualsBadLength {
+
+    int v1;
+    int v2;
+
+    boolean m(String s1) {
+        int l = v2 - v1; // 0 - (-1) = 1. On 64 bit: 0xffffffff00000001
+        char[] arr = new char[l];
+        arr[0] = 'a';
+        String s2 = new String(arr);
+        // The string length is not reloaded but the value computed is
+        // reused so pointer computation must not use
+        // 0xffffffff00000001
+        return s2.equals(s1);
+    }
+
+    // Same thing with String.compareTo()
+    int m2(String s1) {
+        int l = v2 - v1;
+        char[] arr = new char[l+1];
+        arr[0] = 'a';
+        arr[1] = 'b';
+        String s2 = new String(arr);
+        return s2.compareTo(s1);
+    }
+
+    // Same thing with equals() for arrays
+    boolean m3(char[] arr1) {
+        int l = v2 - v1; // 0 - (-1) = 1. On 64 bit: 0xffffffff00000001
+        char[] arr2 = new char[l];
+        arr2[0] = 'a';
+        return Arrays.equals(arr2, arr1);
+    }
+
+    static public void main(String[] args) {
+        TestStringEqualsBadLength tse = new TestStringEqualsBadLength();
+        tse.v1 = -1;
+        tse.v2 = 0;
+        char[] arr = new char[1];
+        arr[0] = 'a';
+        for (int i = 0; i < 20000; i++) {
+            tse.m("a");
+            tse.m2("ab");
+            tse.m3(arr);
+        }
+
+        System.out.println("TEST PASSED");
+    }
+}

test/compiler/regalloc/C1ObjectSpillInLogicOp.java

+/*
+ * Copyright (c) 2013, Oracle and/or its affiliates. All rights reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ */
+
+/*
+ * @test
+ * @bug 8027751
+ * @summary C1 crashes generating G1 post-barrier in Unsafe.getAndSetObject() intrinsic because of the new value spill
+ * @run main/othervm -XX:+UseG1GC C1ObjectSpillInLogicOp
+ *
+ * G1 barriers use logical operators (xor) on T_OBJECT mixed with T_LONG or T_INT.
+ * The current implementation of logical operations on x86 in C1 doesn't allow for long operands to be on stack.
+ * There is a special code in the register allocator that forces long arguments in registers on x86. However T_OBJECT
+ * can be spilled just fine, and in that case the xor emission will fail.
+ */
+
+import java.util.concurrent.atomic.*;
+class C1ObjectSpillInLogicOp {
+  static public void main(String[] args) {
+    AtomicReferenceArray<Integer> x = new AtomicReferenceArray(128);
+    Integer y = new Integer(0);
+    for (int i = 0; i < 50000; i++) {
+      x.getAndSet(i % x.length(), y);
+    }
+  }
+}